Maximem Synap

Maximem Synap is a memory and context management layer for AI agents, shipped as a developer SDK, a REST API, and a hosted MCP endpoint. You send Synap the conversation as it happens, and before your agent replies you ask what is known about this person; Synap returns a short, ranked set of facts, formatted and ready for the prompt. The product is aimed at teams building agents for customer support, sales, voice, healthcare, and multi-agent workflows, so that every conversation does not start from zero. Synap handles entity resolution, temporal reasoning, and multi-level scoping automatically, and the company states there is no vector database to run, no extraction pipeline to build, no retrieval ranker to tune, and no scoping logic to get right - those are the product. Reported accuracy is 92% on LongMemEval and 93.2% on LoCoMo, with in-conversation retrieval under 15ms at P75. Synap exists because agents forget, and the site treats that as an architecture problem rather than an interface nitpick. The listed symptoms are concrete: the agent re-asks for information the user already gave, recommends what the user already rejected, contradicts itself across sessions, or compensates by stuffing everything into the prompt, which produces slow replies and a blown token budget. The company frames these failures as tickets, refunds, and churn. It also argues that every alternative has been tried and each one stops short. A bigger context window runs into U-shaped attention, so models lose the middle; quality falls as the window fills and cost grows quadratically because history is replayed every turn. A vector database and RAG finds text that looks alike but has no view of what is current, no record of where a fact came from, and no idea that 'Acme' and 'Acme Corp' are one company. Agent skills steady a procedure but do not supply a fact - across 528 matched runs, the skill was steadying the procedure 65.7% of the time and supplying a missing fact 4.5% of the time. Summarising as you go compounds losses, and doing it properly means entity extraction, temporal reasoning, contradiction handling and validation, at which point you have built a memory system. Building it in-house is possible, but retrieval is the easy part; the months go into deciding what to keep, resolving entities, handling contradictions, and keeping tenants apart, then maintaining all of it forever. Every one of these workarounds is described as a stand-in for a missing layer. Synap treats memory as three layers rather than one bucket. Short-term memory is the current session, the working memory that most memory tools provide. Long-term memory persists across sessions, per person - the layer a user means when they say an agent remembers them. Organisational memory is shared across users and tenants: policies, product facts, and pricing, meaning company knowledge rather than personal knowledge. The company states that most memory tools give you the session and that the value sits in the two layers above it. Scoping enforces that structure: a request sees its own level and every level above it, never below and never sideways. The out-of-the-box hierarchy is Client, Customer, User - shared product knowledge visible to everyone, tenant-level policies, teams and shared projects, and the facts, preferences and episodes private to one person. One person's memory does not reach another person's session and one tenant's does not reach another tenant's. When three levels is not the right shape, teams can define a custom hierarchy at any depth with the names they already use, such as Hospital, Department, Clinician, Patient. On the read side, Synap focuses on anticipatory retrieval. Context is pre-fetched while the conversation is still going, so it is ready before the agent asks, at under 15ms at P75 in-conversation. That matters most for voice agents, which stay conversational instead of pausing mid-turn. Retrieval is also described as resilient: a bad day for one part of the system is not an outage, and retrieval nets across all stores. Agentic compaction keeps context lean as conversations grow, so cost does not balloon and quality does not rot; it keeps the signal, drops the noise, and tells you when it worked. Validation is part of the pitch, since compaction is reported rather than assumed. Together, these read-side capabilities are what the company presents as the difference between storing memory and actively managing context, and they are the reason a growing conversation does not automatically mean a growing prompt. On the write side, a conversation turn does not land in a database. It runs through a pipeline that turns raw dialogue into structured, scoped memory: asynchronous ingest, extraction of structure rather than raw text, and storage across vector, graph, and file stores - vectors for semantic similarity, a graph for entity relationships, and files for documents and raw material. The write call returns before any of that happens, so it never blocks the agent. What governs the pipeline is a custom context architecture generated for each agent, controlling extraction, scoping, retention and more, rather than a one-size schema. Structured capture and entity resolution mean that 'Sarah', 'Sarah Chen' and 'SC' resolve to one person automatically, and references such as 'my manager' are linked across sessions. Temporal awareness weights recent context higher than stale context. Conscious forgetting processes retractions and contradictions, and because a change never destroys the previous version, every memory can be traced back through provenance. Consolidation runs as background cycles on the stores in three tiers - meditation, a light pass every few hours; nap, a deeper pass once a day; and sleep during quiet hours for deep consolidation and conscious forgetting. Synap's distinctive approach is that it treats memory as an active context-management problem layered on top of storage, rather than a storage problem alone. It works in two calls: you record the conversation as it happens, and before your agent replies you fetch what is known, receiving formatted context ready for the prompt. Everything the pipeline does behind that write call is invisible to the agent - extraction, scoping, retention, and consolidation run behind it. The memory architecture itself is generated for your specific agent instead of being fitted to a universal memory model, which the company contrasts with approaches built on extracted facts plus embeddings or on a temporal knowledge graph. Reads mostly never leave your process because context is pre-fetched while the conversation is still going. The framework is described as native across 23 agent frameworks, with adapters that let developers install the SDK, configure an API key, and start managing context in a few lines of code. The stated outcomes are reliability, cost control, and speed. Agents remember every user across sessions, channels, and months rather than the last twenty turns, so they stop re-asking for information and stop recommending what was already rejected. They remember the organisation through shared policies, product knowledge, and team context made visible to the agents that should see them and isolated from the users and tenants that should not. Context stays lean as conversations grow, which means the token bill does not balloon and quality does not rot; compaction keeps signal over noise and reports when it worked. Context arriving before the agent asks keeps voice agents conversational. Accuracy is reported at 92% on LongMemEval - the benchmark that tests whether a memory system retrieves the right fact from a long conversation and holds that accuracy as the conversation grows - and 93.2% on LoCoMo, with the methodology published and the eval harness open source so teams can run it against any system they are evaluating. Synap is not limited to a fixed list of applications. The site names customer support and sales agents, voice concierges, healthcare assistants, and multi-agent workflows among the places teams run it today, with dedicated use-case pages for healthcare, customer support, sales, voice AI, and multi-agent systems. In support and sales, the relevant value is recall of what a customer said in earlier sessions so an agent does not reopen a resolved issue or contradict a previous answer. In voice, the value is latency: sub-15ms P75 in-conversation retrieval keeps a voice agent conversational instead of pausing to think. In healthcare, the scoping model and custom hierarchy matter - the site illustrates a Hospital, Department, Clinician, Patient shape. In multi-agent workflows, shared organisational context reaches the agents that should see it while per-user and per-tenant memories stay isolated. The primary audience is developers and teams building AI agents who need production-grade memory without building it themselves. Synap provides Python and TypeScript SDKs, a REST API usable from any language, a hosted MCP endpoint for no-code platforms, and a CLI. The documentation states that most developers are up and running in under five minutes. Native integrations cover 23 agent frameworks, including LangChain, LangGraph, LlamaIndex, OpenAI Agents, Pydantic AI, CrewAI, AutoGen, Google ADK, Haystack, Agno, Semantic Kernel, Microsoft Agent Framework, NeMo Agent Toolkit, LiveKit Agents, Pipecat, Claude Agent SDK, Mastra, Vercel AI SDK, Vercel eve, Strands Agents, CAMEL-AI, Smolagents, and deepagents. Security includes encryption in transit and at rest, strict tenant isolation, BYOK for model providers, and on-premise, self-hosted, and air-gapped options, with PII posture applied per kind of data and down to what an individual API key may see. Enterprise plans add VPC and private deployment, SSO and SAML, configurable RBAC, and custom SLAs. The free tier requires no credit card and supports Google or GitHub sign-in, and the SDK plus the benchmark eval harnesses are open source on GitHub. For teams whose agents need to remember, Maximem Synap packages the whole problem - what to keep, how to scope it, when to retrieve it, and when to let it go - into a memory and context layer with reported benchmark accuracy of 92% on LongMemEval and 93.2% on LoCoMo, in-conversation retrieval under 15ms at P75, and native coverage of 23 agent frameworks. It replaces the workarounds teams currently reach for, including larger context windows, vector RAG, agent skills, rolling summaries, or a homegrown system, with a two-call interface and background consolidation, so that every conversation no longer starts from zero.